Domain Transformation-Based Efficient Cost Aggregation for Local Stereo Matching

Abstract

Abstract-Binocular stereo matching is one of the most important algorithms in the field of computer vision. Adaptive support-weight approaches, the current state-of-the-art local methods, produce results comparable to those generated by global methods. However, excessive time consumption is the main problem of these algorithms since the computational complexity is proportionally related to the support window size. In this paper, we present a novel cost aggregation method inspired by domain transformation, a recently proposed dimensionality reduction technique. This transformation enables the aggregation of 2-D cost data to be performed using a sequence of 1-D filters, which lowers computation and memory costs compared to conventional 2-D filters. Experiments show that the proposed method outperforms the state-of-the-art local methods in terms of computational performance, since its computational complexity is independent of the input parameters. Furthermore, according to the experimental results with the Middlebury dataset and realworld images, our algorithm is currently one of the most accurate and efficient local algorithms.

Image-Guided Non-Local Dense Matching With Three-Steps Optimization

“…Large windows take more time for cost aggregation while small windows perform badly in textureless areas. In order to avoid the window definition, the non-local methods, which are based on recursion, were proposed (Yang, 2015;Pham and Jeon, 2013;Cigla and Alantan, 2013;Sun et al, 2014;Cheng et al, 2015), which differed from the local methods in that the cost aggregation of every pixel is supported by the remaining pixels in the whole image for non-local methods. The supports from the remaining pixels depend on the intensity similarity and the cost aggregation path.…”

“…In recent years, several image-guided non-local methods were proposed, of which the basic mathematic models are consistent essentially (Yang, 2015;Pham and Jeon, 2013;Cigla and Alantan, 2013;Sun et al, 2014;Cheng et al, 2015):…”

“…In fact, it is impossible that the intensities of pixels in the same region are exactly the same. In order to strengthen the cost propagation in homogenous intensity regions, traditional methods choose a larger  (  =15~25) to acquire a larger T (Yang, 2015;Pham and Jeon, 2013;Cigla and Alantan, 2013;Sun et al, 2014;Cheng et al, 2015). However, when the cost is propagated between two different regions, T may remain large with a large  , which may bring mismatches in the depth discontinuities.…”

“…A good cost aggregation path should be connected. Most non-local methods define paths through horizontal/vertical scanlines (Pham and Jeon, 2013;Cigla and Alantan, 2013;Sun et al, 2014;Cheng et al, 2015). However, the path, which is described by horizontal/vertical scanlines, may be unconnected even though the beginning and the end belong to the same region, as shown in Figure 3.…”

Global Patch Matching

“…However, they produce fronto-parallel bias in slanted planes, as shown in Figure 1. Figure 1 (Pham and Jeon, 2013), semi-global matching (SGM) (Hirschmuller, 2008), graph cut (GC) (Kolmogorov and Zabih, 2001), image-guided nonlocal dense matching with three-steps optimization (INTS) (Huang, et al, 2016) and stereo matching using non-texture regions and edge information (NTDE) (Kim et al, 2016), respectively. All of above algorithms are 1D label algorithms.…”